In many computational science and engineering applications, the output of a system of interest corresponding to a given input can be queried at different levels of fidelity with different computational footprints. Within online learning and control, multi-fidelity data can be exploited in many stages of the decision-making pipeline, including estimation, exploration, and exploitation. In this talk, we present our research on learning models with multi-fidelity estimation in the context of Reinforcement Learning (RL) and Bayesian Optimization (BO). Specifically, we propose methods for learning a Q-function (in RL) and a surrogate model (in BO) that can exploit the cross-correlation among multi-fidelity outputs to improve the predictive capability of the model. The efficacy of our proposed multi-fidelity learning techniques are theoretically analyzed in RL, and empirically demonstrated in BO applied to Neural Architecture Search (NAS).

Zoom Link:  https://argonne.zoomgov.com/j/1611891959

See all upcoming talks at http://wordpress.cels.anl.gov/lans-seminars/